Probabilistic Tissue Mapping for Tumor Segmentation and Infiltration Detection of Glioma

Abstract

Segmentation of glioma structures is vital for therapy planning. Although state of the art algorithms achieve impressive results when compared to ground-truth manual delineations, one could argue that the binary nature of these labels does not properly reflect the underlying biology, nor does it account for uncertainties in the predicted segmentations. Moreover, the tumor infiltration beyond the contrast-enhanced lesion - visually imperceptible on imaging - is often ignored despite its potential role in tumor recurrence. We propose an intensity-based probabilistic model for brain tissue mapping based on conventional MRI sequences. We evaluated its value in the binary segmentation of the tumor and its subregions, and in the visualisation of possible infiltration. The model achieves a median Dice of 0.82 in the detection of the whole tumor, but suffers from confusion between different subregions. Preliminary results for the tumor probability maps encourage further investigation of the model regarding infiltration detection.